A Detection Method for Seeding Temperature in Czochralski Silicon Crystal Growth Based on Multi-Sensor Data Fusion

Area of Science:

• Materials Science and Engineering
• Semiconductor Manufacturing

Background:

• The Czochralski method is crucial for producing power-electronics-grade silicon crystals.
• Accurate control of the solid-liquid interface temperature during the seeding stage is vital for crystal quality and yield.
• Direct measurement of the solid-liquid interface temperature is not feasible, and current manual methods lack quantitative assessment.

Purpose of the Study:

• To develop a quantitative method for evaluating the seeding stage temperature in silicon crystal growth.
• To address the limitations of direct temperature measurement and manual visual inspection.
• To improve crystal quality and production yield by ensuring appropriate initial seeding temperatures.

Main Methods:

• Proposed using the seeding stage duration as a proxy for temperature evaluation.
• Developed an improved multimodal fusion regression network integrating pyrometer data and meniscus features.
• Employed wavelet transform for time-frequency analysis, multimodal feature fusion (MFF), channel attention (CA), and spatial attention (SA), with a random vector functional link network (RVFLN) for prediction.

Main Results:

• Successfully established an indirect quantitative relationship between multi-sensor data and the seeding temperature.
• The proposed model demonstrated superior detection performance compared to previous methods.
• Feature extraction strategies were validated, confirming their effectiveness.

Conclusions:

• The developed multimodal fusion network provides an effective indirect method for quantifying the solid-liquid interface temperature during silicon crystal seeding.
• This approach enhances the control over initial seeding temperature, leading to potential improvements in semiconductor crystal quality and production yield.
• The study highlights the efficacy of using time-frequency representations and attention mechanisms for complex sensor data fusion in materials processing.